AI in Legal Research and Case Law Search

Legal research has run on computers for decades. Westlaw and LexisNexis built keyword and Boolean search over case law starting in the 1970s. The newer generation of tools adds two capabilities. The first is semantic search, which turns queries and passages into numerical vectors (embeddings), so a tool can find a relevant opinion even when it uses different words than the query. The second is generative models, which read the retrieved passages and draft an answer or a memo. Major products include Thomson Reuters' CoCounsel (built on Casetext, which Thomson Reuters acquired in 2023), AI features in Westlaw and Lexis+ AI, and Harvey, which is popular with large firms. Most legal-specific tools use retrieval-augmented generation. They search a curated database of real authorities first, then ask the model to answer from what was retrieved, with citations linking back to the sources. The main risk is hallucination. By default, a general chatbot does not look anything up. It predicts plausible text, and a plausible legal citation has a case name, reporter volume and page number whether or not the case exists. In Mata v. Avianca (2023), attorneys filed a brief citing several cases that ChatGPT had invented, and they stood by those cases when the court questioned them. Judge P. Kevin Castel imposed a $5,000 sanction. Courts have reported many similar incidents since, and some judges now require lawyers to disclose or certify how they used generative AI. A common misconception is that retrieval solves the problem. It reduces fabricated citations, but a tool can still cite a real case for a point it does not support, misstate a holding, or miss that a case was overruled. A 2024 study by Stanford researchers found that leading legal research tools still gave incorrect or misgrounded answers to a meaningful share of test questions, though less often than general-purpose chatbots. Professional rules put the responsibility on the lawyer who signs the filing, not on the software.